An Introduction to Structured P2P Overlay Networks With Cakes

1
An Introduction to Structured P2P Overlay
Networks (With Cakes)

• Rob Minson
• Rich Price
• Tien Tuan Anh Dinh

2
Overview

• Peer-to-Peer overlay technology is mature and
  widely deployed.
• Largely unstructured, used in non-critical
  (usually nefarious) environments
• Since 2000 much research in to turning P2P
  overlays in to powerful distributed computing
  platforms.

3
Unstructured Networks

• Example Application Network File System
• Strings (home/x) map to data
• All data located at globally known IP

4
Unstructured Networks

• Unstructured P2P implementation
• Each node knows the IPs of a random subset of
  others

5
Unstructured Networks

• Searching an unstructured network
• Flooding each node forwards to all neighbours.
• Guaranteed to locate data if it exists
• Very inefficient for non-trivial networks

6
Unstructured Networks
7
Unstructured Networks

• Searching an unstructured network
• Constrained flooding each query only traverses a
  distance from its origin equal to its
  time-to-live value (TTL).
• Reduces bandwidth forwarding overhead
• Only locates data with probabilistic guarantees

8
Unstructured Networks
9
Unstructured Networks
10
Adding Structure

• Key Points for this section
• Super node network has very high degree
• High maintenance cost (join/leave)
• Low latency traversals
• Standard node network has very low degree
• Low maintenance cost
• High latency traversals
• This is essential trade-off

11
Adding Structure
12
Structured Networks

• Basic motivations
• Deterministic searches (flooding)
• Low bandwidth usage (constrained flooding)
• Low latency (high-degree)
• Low maintenance overhead (low-degree)

13
Structured Networks

• Chord
• Globally known low-collision hash function
• IP addresses and data items both hashed using
  same function in to a single address space

/home/rzm/mp3s
788
1000
0
1024
348
/home/ttd/donuts
147.188.43.53
14
Structured Networks

• Chord
• Address space is arranged in to a ring
• Nodes are attached to their neighbours

0
1000
114
609
15
Structured Networks

• Searching in Chord
• Each node stores the data between itself and its
  predecessor in the ring.

0
1000
114
/home/ttd/donuts
348
788
/home/rzm/mp3s
609
16
Structured Networks

• Searching in Chord
• Long-distance links (fingers) are added to each
  node at logarithmically increasing distances
  (allows a binary search)

target
peer
1000
114 1
609
114
114 2
609
114 4
609
114 8
609
114 16
609
114 32
609
114 64
609
114 128
609
114 256
609
114 512
1000
609
114 1024
self
17
Structured Networks

• Searching in Chord
• Long-distance links (fingers) are added to each
  node at logarithmically increasing distances
  (allows a binary search)

target
peer
1000
114 1
296
114
114 2
296
114 4
296
114 8
296
114 16
296
114 32
296
114 64
296
296
114 128
296
114 256
609
114 512
1000
609
114 1024
self
18
Structured Networks

• Searching in Chord
• Nodes forward queries to the finger pointer with
  the highest address less than the target address
  (eg. lookup(744))

1000
Q
114
296
775
744
609
498
19
Structured Networks

• Searching in Chord
• Nodes forward queries to the finger pointer with
  the highest address less than the target address
  (eg. lookup(744))

1000
114
296
775
744
Q
609
498
20
Structured Networks

• Searching in Chord
• Nodes forward queries to the finger pointer with
  the highest address less than the target address
  (eg. lookup(744))
• Unless the target is between you and your
  successor!

1000
114
296
775
744
609
498
Q
21
Structured Networks

• Several Structured Networks exist
• CAN (2000)
• N-Dimensional toroidal address space
• Pastry (2001)
• Hybrid tree/ring routing IP latency estimation
• Kademlia (2002)
• Widely deployed (eg. Azeurus, BitTorrent
  Mainline)
• All use essentially the same ideas

22
Structured Networks

• Structured network essentials
• Deterministic
• Unique, deterministic responsibility for data
• Low Bandwidth
• No route-branching
• Low Latency
• O(logn)-hop routing
• Low Maintenance
• But significantly higher than unstructured
  networks

23
Nodes joining (1)

• Picks a bootstrapping node
• Initializes successor,
• predecessor and fingers

0
1000
114
/home/ttd/donuts
788
348
/home/rzm/mp3s
Search using 144 Successor key 4501 -gt Node
609 Predecessor successors predecessor -gt Node
114 Finger2 key 4502 -gt Node 609 Finger9
key 450256-gt Node 1000 Finger10 key 450512
-gt Node 1000
450
500
609
/home/rzm/makeup
24
Nodes joining (2)

• Update other nodes
• Transfer keys from successor node

Node 114s changes Successor Node 609 -gt
450 Finger2 to Finger9 Node 609 -gt 450
0
Node 1000s changes Finger9 Node 609 -gt 450
1000
114
/home/ttd/donuts
788
348
/home/rzm/mp3s
450
500
609
/home/rzm/makeup
Node 609s changes Predecessor Node 114 -gt 450
25
Nodes leaving

• Leaving in Chord (gracefully)
• Update other nodes
• Move keys to successor nodes

Node 114s changes Successor Node 609 -gt
450 Finger2 to Finger9 Node 609 -gt 450
0
1000
Node 1000s changes Predecessor Node 609 -gt
450 Finger10 Node 609 -gt 1000
114
/home/ttd/donuts
788
348
/home/rzm/mp3s
450
500
Node 450s changes Successor Node 609 -gt
1000 Finger2 to Finger8 Node 609 -gt 1000
609
/home/rzm/makeup
26
Concurrent joins/leaves

• Frequent join/leave (high churn rate)
• Maintenance protocol
• Can overload the network
• Can jeopardise deterministic routing
• Longer routing path
• Query failure

A
B
C
27
Open Research Topics

• Complex queries
• Load balancing
• Maintenance under churn
• Security
• Applications
• Application-Layer Multicast
• Distributed Storage
• P2P Gaming
• Streaming Media
• etc

28
Overlay Research at cs.bham

• Simulation (rmp/ttd)
• Security (ttd)
• Advanced Search (rzm)
• Robustness (rmp)
• etc?

29
Thanks!

• (cakes may have contained nuts)